1 /*****************************************************************************
2 * mc.c: h264 encoder library (Motion Compensation)
3 *****************************************************************************
4 * Copyright (C) 2003-2008 x264 project
6 * Authors: Laurent Aimar <fenrir@via.ecp.fr>
7 * Loren Merritt <lorenm@u.washington.edu>
9 * This program is free software; you can redistribute it and/or modify
10 * it under the terms of the GNU General Public License as published by
11 * the Free Software Foundation; either version 2 of the License, or
12 * (at your option) any later version.
14 * This program is distributed in the hope that it will be useful,
15 * but WITHOUT ANY WARRANTY; without even the implied warranty of
16 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
17 * GNU General Public License for more details.
19 * You should have received a copy of the GNU General Public License
20 * along with this program; if not, write to the Free Software
21 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02111, USA.
22 *****************************************************************************/
37 static inline void pixel_avg( uint8_t *dst, int i_dst_stride,
38 uint8_t *src1, int i_src1_stride,
39 uint8_t *src2, int i_src2_stride,
40 int i_width, int i_height )
42 for( int y = 0; y < i_height; y++ )
44 for( int x = 0; x < i_width; x++ )
45 dst[x] = ( src1[x] + src2[x] + 1 ) >> 1;
47 src1 += i_src1_stride;
48 src2 += i_src2_stride;
52 static inline void pixel_avg_wxh( uint8_t *dst, int i_dst, uint8_t *src1, int i_src1, uint8_t *src2, int i_src2, int width, int height )
54 for( int y = 0; y < height; y++ )
56 for( int x = 0; x < width; x++ )
57 dst[x] = ( src1[x] + src2[x] + 1 ) >> 1;
64 /* Implicit weighted bipred only:
65 * assumes log2_denom = 5, offset = 0, weight1 + weight2 = 64 */
66 #define op_scale2(x) dst[x] = x264_clip_uint8( (src1[x]*i_weight1 + src2[x]*i_weight2 + (1<<5)) >> 6 )
67 static inline void pixel_avg_weight_wxh( uint8_t *dst, int i_dst, uint8_t *src1, int i_src1, uint8_t *src2, int i_src2, int width, int height, int i_weight1 )
69 const int i_weight2 = 64 - i_weight1;
70 for( int y = 0; y<height; y++, dst += i_dst, src1 += i_src1, src2 += i_src2 )
74 if(width==2) continue;
77 if(width==4) continue;
82 if(width==8) continue;
95 #define PIXEL_AVG_C( name, width, height ) \
96 static void name( uint8_t *pix1, int i_stride_pix1, \
97 uint8_t *pix2, int i_stride_pix2, \
98 uint8_t *pix3, int i_stride_pix3, int weight ) \
101 pixel_avg_wxh( pix1, i_stride_pix1, pix2, i_stride_pix2, pix3, i_stride_pix3, width, height ); \
103 pixel_avg_weight_wxh( pix1, i_stride_pix1, pix2, i_stride_pix2, pix3, i_stride_pix3, width, height, weight ); \
105 PIXEL_AVG_C( pixel_avg_16x16, 16, 16 )
106 PIXEL_AVG_C( pixel_avg_16x8, 16, 8 )
107 PIXEL_AVG_C( pixel_avg_8x16, 8, 16 )
108 PIXEL_AVG_C( pixel_avg_8x8, 8, 8 )
109 PIXEL_AVG_C( pixel_avg_8x4, 8, 4 )
110 PIXEL_AVG_C( pixel_avg_4x8, 4, 8 )
111 PIXEL_AVG_C( pixel_avg_4x4, 4, 4 )
112 PIXEL_AVG_C( pixel_avg_4x2, 4, 2 )
113 PIXEL_AVG_C( pixel_avg_2x4, 2, 4 )
114 PIXEL_AVG_C( pixel_avg_2x2, 2, 2 )
116 static void x264_weight_cache( x264_t *h, x264_weight_t *w )
118 w->weightfn = h->mc.weight;
120 #define opscale(x) dst[x] = x264_clip_uint8( ((src[x] * weight->i_scale + (1<<(weight->i_denom - 1))) >> weight->i_denom) + weight->i_offset )
121 #define opscale_noden(x) dst[x] = x264_clip_uint8( src[x] * weight->i_scale + weight->i_offset )
122 static inline void mc_weight( uint8_t *dst, int i_dst_stride, uint8_t *src, int i_src_stride, const x264_weight_t *weight, int i_width, int i_height )
124 if( weight->i_denom >= 1 )
126 for( int y = 0; y < i_height; y++, dst += i_dst_stride, src += i_src_stride )
127 for( int x = 0; x < i_width; x++ )
132 for( int y = 0; y < i_height; y++, dst += i_dst_stride, src += i_src_stride )
133 for( int x = 0; x < i_width; x++ )
138 #define MC_WEIGHT_C( name, lx ) \
139 static void name( uint8_t *dst, int i_dst_stride, uint8_t *src, int i_src_stride, const x264_weight_t *weight, int height ) \
141 if( weight->i_denom >= 1 ) \
143 for( int y = 0; y < height; y++, dst += i_dst_stride, src += i_src_stride ) \
144 for( int x = 0; x < lx; x++ ) \
149 for( int y = 0; y < height; y++, dst += i_dst_stride, src += i_src_stride ) \
150 for( int x = 0; x < lx; x++ ) \
151 opscale_noden( x ); \
155 MC_WEIGHT_C( mc_weight_w20, 20 )
156 MC_WEIGHT_C( mc_weight_w16, 16 )
157 MC_WEIGHT_C( mc_weight_w12, 12 )
158 MC_WEIGHT_C( mc_weight_w8, 8 )
159 MC_WEIGHT_C( mc_weight_w4, 4 )
160 MC_WEIGHT_C( mc_weight_w2, 2 )
162 static weight_fn_t x264_mc_weight_wtab[6] =
171 const x264_weight_t weight_none[3] = { {{0}} };
172 static void mc_copy( uint8_t *src, int i_src_stride, uint8_t *dst, int i_dst_stride, int i_width, int i_height )
174 for( int y = 0; y < i_height; y++ )
176 memcpy( dst, src, i_width );
183 #define TAPFILTER(pix, d) ((pix)[x-2*d] + (pix)[x+3*d] - 5*((pix)[x-d] + (pix)[x+2*d]) + 20*((pix)[x] + (pix)[x+d]))
184 static void hpel_filter( uint8_t *dsth, uint8_t *dstv, uint8_t *dstc, uint8_t *src,
185 int stride, int width, int height, int16_t *buf )
187 for( int y = 0; y < height; y++ )
189 for( int x = -2; x < width+3; x++ )
191 int v = TAPFILTER(src,stride);
192 dstv[x] = x264_clip_uint8( (v + 16) >> 5 );
195 for( int x = 0; x < width; x++ )
196 dstc[x] = x264_clip_uint8( (TAPFILTER(buf+2,1) + 512) >> 10 );
197 for( int x = 0; x < width; x++ )
198 dsth[x] = x264_clip_uint8( (TAPFILTER(src,1) + 16) >> 5 );
206 static const uint8_t hpel_ref0[16] = {0,1,1,1,0,1,1,1,2,3,3,3,0,1,1,1};
207 static const uint8_t hpel_ref1[16] = {0,0,0,0,2,2,3,2,2,2,3,2,2,2,3,2};
209 static void mc_luma( uint8_t *dst, int i_dst_stride,
210 uint8_t *src[4], int i_src_stride,
212 int i_width, int i_height, const x264_weight_t *weight )
214 int qpel_idx = ((mvy&3)<<2) + (mvx&3);
215 int offset = (mvy>>2)*i_src_stride + (mvx>>2);
216 uint8_t *src1 = src[hpel_ref0[qpel_idx]] + offset + ((mvy&3) == 3) * i_src_stride;
218 if( qpel_idx & 5 ) /* qpel interpolation needed */
220 uint8_t *src2 = src[hpel_ref1[qpel_idx]] + offset + ((mvx&3) == 3);
221 pixel_avg( dst, i_dst_stride, src1, i_src_stride,
222 src2, i_src_stride, i_width, i_height );
223 if( weight->weightfn )
224 mc_weight( dst, i_dst_stride, dst, i_dst_stride, weight, i_width, i_height );
226 else if( weight->weightfn )
227 mc_weight( dst, i_dst_stride, src1, i_src_stride, weight, i_width, i_height );
229 mc_copy( src1, i_src_stride, dst, i_dst_stride, i_width, i_height );
232 static uint8_t *get_ref( uint8_t *dst, int *i_dst_stride,
233 uint8_t *src[4], int i_src_stride,
235 int i_width, int i_height, const x264_weight_t *weight )
237 int qpel_idx = ((mvy&3)<<2) + (mvx&3);
238 int offset = (mvy>>2)*i_src_stride + (mvx>>2);
239 uint8_t *src1 = src[hpel_ref0[qpel_idx]] + offset + ((mvy&3) == 3) * i_src_stride;
241 if( qpel_idx & 5 ) /* qpel interpolation needed */
243 uint8_t *src2 = src[hpel_ref1[qpel_idx]] + offset + ((mvx&3) == 3);
244 pixel_avg( dst, *i_dst_stride, src1, i_src_stride,
245 src2, i_src_stride, i_width, i_height );
246 if( weight->weightfn )
247 mc_weight( dst, *i_dst_stride, dst, *i_dst_stride, weight, i_width, i_height );
250 else if( weight->weightfn )
252 mc_weight( dst, *i_dst_stride, src1, i_src_stride, weight, i_width, i_height );
257 *i_dst_stride = i_src_stride;
262 /* full chroma mc (ie until 1/8 pixel)*/
263 static void mc_chroma( uint8_t *dst, int i_dst_stride,
264 uint8_t *src, int i_src_stride,
266 int i_width, int i_height )
272 int cA = (8-d8x)*(8-d8y);
273 int cB = d8x *(8-d8y);
274 int cC = (8-d8x)*d8y;
277 src += (mvy >> 3) * i_src_stride + (mvx >> 3);
278 srcp = &src[i_src_stride];
280 for( int y = 0; y < i_height; y++ )
282 for( int x = 0; x < i_width; x++ )
283 dst[x] = ( cA*src[x] + cB*src[x+1] + cC*srcp[x] + cD*srcp[x+1] + 32 ) >> 6;
286 srcp += i_src_stride;
291 static void mc_copy_w##W( uint8_t *dst, int i_dst, uint8_t *src, int i_src, int i_height ) \
293 mc_copy( src, i_src, dst, i_dst, W, i_height ); \
299 void x264_plane_copy_c( uint8_t *dst, int i_dst,
300 uint8_t *src, int i_src, int w, int h)
304 memcpy( dst, src, w );
310 static void prefetch_fenc_null( uint8_t *pix_y, int stride_y,
311 uint8_t *pix_uv, int stride_uv, int mb_x )
314 static void prefetch_ref_null( uint8_t *pix, int stride, int parity )
317 static void memzero_aligned( void * dst, int n )
322 static void integral_init4h( uint16_t *sum, uint8_t *pix, int stride )
324 int v = pix[0]+pix[1]+pix[2]+pix[3];
325 for( int x = 0; x < stride-4; x++ )
327 sum[x] = v + sum[x-stride];
328 v += pix[x+4] - pix[x];
332 static void integral_init8h( uint16_t *sum, uint8_t *pix, int stride )
334 int v = pix[0]+pix[1]+pix[2]+pix[3]+pix[4]+pix[5]+pix[6]+pix[7];
335 for( int x = 0; x < stride-8; x++ )
337 sum[x] = v + sum[x-stride];
338 v += pix[x+8] - pix[x];
342 static void integral_init4v( uint16_t *sum8, uint16_t *sum4, int stride )
344 for( int x = 0; x < stride-8; x++ )
345 sum4[x] = sum8[x+4*stride] - sum8[x];
346 for( int x = 0; x < stride-8; x++ )
347 sum8[x] = sum8[x+8*stride] + sum8[x+8*stride+4] - sum8[x] - sum8[x+4];
350 static void integral_init8v( uint16_t *sum8, int stride )
352 for( int x = 0; x < stride-8; x++ )
353 sum8[x] = sum8[x+8*stride] - sum8[x];
356 void x264_frame_init_lowres( x264_t *h, x264_frame_t *frame )
358 uint8_t *src = frame->plane[0];
359 int i_stride = frame->i_stride[0];
360 int i_height = frame->i_lines[0];
361 int i_width = frame->i_width[0];
363 // duplicate last row and column so that their interpolation doesn't have to be special-cased
364 for( int y = 0; y < i_height; y++ )
365 src[i_width+y*i_stride] = src[i_width-1+y*i_stride];
366 memcpy( src+i_stride*i_height, src+i_stride*(i_height-1), i_width+1 );
367 h->mc.frame_init_lowres_core( src, frame->lowres[0], frame->lowres[1], frame->lowres[2], frame->lowres[3],
368 i_stride, frame->i_stride_lowres, frame->i_width_lowres, frame->i_lines_lowres );
369 x264_frame_expand_border_lowres( frame );
371 memset( frame->i_cost_est, -1, sizeof(frame->i_cost_est) );
373 for( int y = 0; y < h->param.i_bframe + 2; y++ )
374 for( int x = 0; x < h->param.i_bframe + 2; x++ )
375 frame->i_row_satds[y][x][0] = -1;
377 for( int y = 0; y <= !!h->param.i_bframe; y++ )
378 for( int x = 0; x <= h->param.i_bframe; x++ )
379 frame->lowres_mvs[y][x][0][0] = 0x7FFF;
382 static void frame_init_lowres_core( uint8_t *src0, uint8_t *dst0, uint8_t *dsth, uint8_t *dstv, uint8_t *dstc,
383 int src_stride, int dst_stride, int width, int height )
385 for( int y = 0; y < height; y++ )
387 uint8_t *src1 = src0+src_stride;
388 uint8_t *src2 = src1+src_stride;
389 for( int x = 0; x<width; x++ )
391 // slower than naive bilinear, but matches asm
392 #define FILTER(a,b,c,d) ((((a+b+1)>>1)+((c+d+1)>>1)+1)>>1)
393 dst0[x] = FILTER(src0[2*x ], src1[2*x ], src0[2*x+1], src1[2*x+1]);
394 dsth[x] = FILTER(src0[2*x+1], src1[2*x+1], src0[2*x+2], src1[2*x+2]);
395 dstv[x] = FILTER(src1[2*x ], src2[2*x ], src1[2*x+1], src2[2*x+1]);
396 dstc[x] = FILTER(src1[2*x+1], src2[2*x+1], src1[2*x+2], src2[2*x+2]);
399 src0 += src_stride*2;
407 #if defined(__GNUC__) && (defined(ARCH_X86) || defined(ARCH_X86_64))
408 // gcc isn't smart enough to use the "idiv" instruction
409 static ALWAYS_INLINE int32_t div_64_32(int64_t x, int32_t y)
411 int32_t quotient, remainder;
413 :"=a"(quotient), "=d"(remainder)
414 :"a"((uint32_t)x), "d"((int32_t)(x>>32)), "r"(y)
419 #define div_64_32(x,y) ((x)/(y))
422 /* Estimate the total amount of influence on future quality that could be had if we
423 * were to improve the reference samples used to inter predict any given macroblock. */
424 static void mbtree_propagate_cost( int *dst, uint16_t *propagate_in, uint16_t *intra_costs,
425 uint16_t *inter_costs, uint16_t *inv_qscales, int len )
427 for( int i = 0; i < len; i++ )
429 int propagate_amount = propagate_in[i] + ((intra_costs[i] * inv_qscales[i] + 128)>>8);
430 dst[i] = div_64_32((int64_t)propagate_amount * (intra_costs[i] - (inter_costs[i] & LOWRES_COST_MASK)), intra_costs[i]);
434 void x264_mc_init( int cpu, x264_mc_functions_t *pf )
436 pf->mc_luma = mc_luma;
437 pf->get_ref = get_ref;
438 pf->mc_chroma = mc_chroma;
440 pf->avg[PIXEL_16x16]= pixel_avg_16x16;
441 pf->avg[PIXEL_16x8] = pixel_avg_16x8;
442 pf->avg[PIXEL_8x16] = pixel_avg_8x16;
443 pf->avg[PIXEL_8x8] = pixel_avg_8x8;
444 pf->avg[PIXEL_8x4] = pixel_avg_8x4;
445 pf->avg[PIXEL_4x8] = pixel_avg_4x8;
446 pf->avg[PIXEL_4x4] = pixel_avg_4x4;
447 pf->avg[PIXEL_4x2] = pixel_avg_4x2;
448 pf->avg[PIXEL_2x4] = pixel_avg_2x4;
449 pf->avg[PIXEL_2x2] = pixel_avg_2x2;
451 pf->weight = x264_mc_weight_wtab;
452 pf->offsetadd = x264_mc_weight_wtab;
453 pf->offsetsub = x264_mc_weight_wtab;
454 pf->weight_cache = x264_weight_cache;
456 pf->copy_16x16_unaligned = mc_copy_w16;
457 pf->copy[PIXEL_16x16] = mc_copy_w16;
458 pf->copy[PIXEL_8x8] = mc_copy_w8;
459 pf->copy[PIXEL_4x4] = mc_copy_w4;
461 pf->plane_copy = x264_plane_copy_c;
462 pf->hpel_filter = hpel_filter;
464 pf->prefetch_fenc = prefetch_fenc_null;
465 pf->prefetch_ref = prefetch_ref_null;
466 pf->memcpy_aligned = memcpy;
467 pf->memzero_aligned = memzero_aligned;
468 pf->frame_init_lowres_core = frame_init_lowres_core;
470 pf->integral_init4h = integral_init4h;
471 pf->integral_init8h = integral_init8h;
472 pf->integral_init4v = integral_init4v;
473 pf->integral_init8v = integral_init8v;
475 pf->mbtree_propagate_cost = mbtree_propagate_cost;
478 x264_mc_init_mmx( cpu, pf );
481 if( cpu&X264_CPU_ALTIVEC )
482 x264_mc_altivec_init( pf );
485 x264_mc_init_arm( cpu, pf );
489 void x264_frame_filter( x264_t *h, x264_frame_t *frame, int mb_y, int b_end )
491 const int b_interlaced = h->sh.b_mbaff;
492 const int stride = frame->i_stride[0] << b_interlaced;
493 const int width = frame->i_width[0];
494 int start = (mb_y*16 >> b_interlaced) - 8; // buffer = 4 for deblock + 3 for 6tap, rounded to 8
495 int height = ((b_end ? frame->i_lines[0] : mb_y*16) >> b_interlaced) + 8;
496 int offs = start*stride - 8; // buffer = 3 for 6tap, aligned to 8 for simd
498 if( mb_y & b_interlaced )
501 for( int y = 0; y <= b_interlaced; y++, offs += frame->i_stride[0] )
504 frame->filtered[1] + offs,
505 frame->filtered[2] + offs,
506 frame->filtered[3] + offs,
507 frame->plane[0] + offs,
508 stride, width + 16, height - start,
512 /* generate integral image:
513 * frame->integral contains 2 planes. in the upper plane, each element is
514 * the sum of an 8x8 pixel region with top-left corner on that point.
515 * in the lower plane, 4x4 sums (needed only with --partitions p4x4). */
517 if( frame->integral )
521 memset( frame->integral - PADV * stride - PADH, 0, stride * sizeof(uint16_t) );
526 for( int y = start; y < height; y++ )
528 uint8_t *pix = frame->plane[0] + y * stride - PADH;
529 uint16_t *sum8 = frame->integral + (y+1) * stride - PADH;
531 if( h->frames.b_have_sub8x8_esa )
533 h->mc.integral_init4h( sum8, pix, stride );
535 sum4 = sum8 + stride * (frame->i_lines[0] + PADV*2);
537 h->mc.integral_init4v( sum8, sum4, stride );
541 h->mc.integral_init8h( sum8, pix, stride );
543 h->mc.integral_init8v( sum8-8*stride, stride );